opengate-grounding

Deterministic grounding check for evidence-grounded AI — no LLM judge.

Is an AI answer actually supported by the context it was given? opengate-grounding answers that with pure, reproducible logic — the Python port of OpenGATE's check_grounding. It catches the three grounding failures that matter in production:

• Dropped facts — the specific facts a correct answer must contain actually appear.
• Fabricated numbers — every figure in the answer traces back to the context or the question.
• Failure to abstain — when the context can't answer, the model must decline rather than invent.

Because it's deterministic (no grader model, no API key), it's free, instant, and safe to run on every answer — inline in an agent or as a CI gate. It's the same check that powers the OpenGATE evaluation framework and the OpenGATE MCP server.

Install

pip install opengate-grounding

No dependencies. Python 3.8+.

Quick start

from opengate_grounding import check_grounding

result = check_grounding(
    answer="You have 30 days to request a refund, with no restocking fee.",
    context="Customers may request a full refund within 30 days. There is no restocking fee.",
    anchors=["30 days", "no restocking fee"],
)

result.grounded            # True
result.issues              # []

A failing answer tells you exactly why:

r = check_grounding(
    answer="30 days, no restocking fee, plus a 90-day extension.",
    context="Customers may request a full refund within 30 days. There is no restocking fee.",
    anchors=["30 days", "no restocking fee"],
)
r.grounded             # False
r.ungrounded_numbers   # ['90']
r.issues               # ['ungrounded number "90" — not in the provided context']

Unanswerable questions must abstain:

r = check_grounding(
    answer="That isn't in the provided context — contact sales.",
    context="Pro plan has a 30-day trial.",
    answerable=False,
)
r.grounded    # True
r.abstained   # True

Gate it in your tests

from opengate_grounding import assert_grounded

def test_refund_answer():
    assert_grounded(
        answer=my_rag("How long is the refund window?"),
        context=retrieved_passages,
        anchors=["30 days", "no restocking fee"],
    )

On failure the assertion names the concrete issues, so a grounding regression shows up as a red build.

Use it as a DeepEval metric

pip install "opengate-grounding[deepeval]"

from deepeval import evaluate
from deepeval.test_case import LLMTestCase
from opengate_grounding.deepeval_metric import GroundingMetric

case = LLMTestCase(
    input="How long is the refund window?",
    actual_output="You have 30 days, with no restocking fee.",
    retrieval_context=["Refund within 30 days. No restocking fee."],
    additional_metadata={"anchors": ["30 days", "no restocking fee"]},
)

evaluate([case], [GroundingMetric()])

GroundingMetric maps actual_output → answer, retrieval_context → context, and input → question. Per-case controls (anchors, answerable, allowed_new_numbers) go on additional_metadata or the metric constructor. Score is 1.0 when grounded, 0.0 otherwise; with the default threshold=1.0 a case passes only when fully grounded. Pair it with DeepEval's model-graded metrics for breadth, and use this one to gate the grounding.

API

check_grounding(answer, context, *, question=None, anchors=None, allowed_new_numbers=None, answerable=True, abstain_markers=None) -> GroundingResult

GroundingResult — grounded: bool, anchors_missed: list[str], ungrounded_numbers: list[str], abstained: bool, issues: list[str]. Truthy when grounded (so if check_grounding(...): works).

Also exported: Anchor, contains, numbers_in, flatten_context, DEFAULT_ABSTAIN.

What it deliberately won't do

• No LLM-as-judge. Scores are deterministic checks — reproducible, free, and your judgment lives in the anchors you supply, not a grader model's.
• No general "quality" score. It measures grounding, not fluency or helpfulness.
• It never sees your model. It inspects the answer text you give it — no weights, no internals.

License

MIT © PharmaTools.AI. Part of the OpenGATE project.